Bounded logit attention

Implementation of the main experiment from https://arxiv.org/abs/2105.14824

Example usage

Train a model with a bounded logit attention module for the Caltech birds 2011 dataset.
python3 run.py --dataset caltech_birds2011 --preset BLA

Weights of trained models are saved in weights_dir/ and logs of loss, accuracy, and top-5 accuracy for the uninterpretable baseline model, the model using the soft training-time explanations, and the model using the hard test-time explanations are saved in logs/.

Usage

Dataset

-d,--dataset
Name of the dataset to use, options are {cats_vs_dogs, stanford_dogs, caltech_birds2011}

Preset

-s,--preset
Preset configurations (see paper for details) {L2X-F , BLA, BLA-T, BLA-PH}

• L2X-F: learning to explain on feature level (fixed size explanations)
• BLA: bounded logit attention (without thresholding)
• BLA-T: bounded logit attention with thresolding
• BLA-PH: post-hoc setup (head is frozen and only the explanation module is trained)

Options

(Overwritten by presets where applicable.)

-f, --fixed-size
Produce fixed size explanations (i.e. learning to explain on feature level, not bounded logit attention)

-t, --thresholding
Thresholding using gamma = 0.02.

-p, --post-hoc
Post-hoc setup, i.e. freeze head during training of explanation module

-r, --force-retraining
Retrain, rather than using saved weights from weights_dir/